2,048 NVIDIA HGX B300 GPUs — eight per node, air-cooled — in a live Class A datacenter in South Florida, on Lenovo ThinkSystem SR680a V4 servers, delivered in two phases: 184 nodes (1,472 GPUs) online in February 2027, plus a further 72 nodes (576 GPUs) with timing to be confirmed. Reserved capacity is priced compute-only at $4.60/GPU-hr on a 5-year term or $5.30/GPU-hr on a 3-year term, with a 20% deposit for investment-grade customers and 25% for others.
The site has 5.5 MVA of utility power energized today with no utility queue, which is what makes a February 2027 Phase 1 target credible: the GPUs are the gating item, not the power. Air-cooled HGX B300 on standard 8-GPU nodes suits dense training and large-batch inference without a liquid-cooling dependency. The public tile shows the headline specification; the complete data room — node and network specs, exact location, counterparty and full term sheet — is shared once both sides agree to connect.
Cluster size, site and counterparty details are confirmed by Ax3 before any introduction.
